The Gemini Developer API still has a real free tier as of August 15, 2026. A new developer can create a project and API key in Google AI Studio, choose an eligible model, and send an API request without first moving the project to paid service.
That answer is accurate but incomplete. “Free” can mean that a model has a zero-dollar input and output row, that your project currently has an active quota, or that Google AI Studio lets you test a model. Those are separate conditions. A usable decision also depends on region, data handling, and how much failure your application can tolerate.
The four gates an unpaid project must pass
Use this order before you spend time integrating:
- Region: Google AI Studio and the Gemini API must be available where you access the service and where the relevant terms permit your client to operate.
- Model price: the exact model ID must show a free input and output path on Google's current pricing page.
- Project limits: AI Studio must show useful active RPM, TPM, and RPD for the project that owns the key.
- Workload fit: the data terms and variable capacity of unpaid service must be acceptable for the job.
Passing one gate does not waive the others. A model can show free token pricing while your project has little usable headroom. A generous-looking project quota is irrelevant if the input is confidential. An API key is not a separate allowance: keys inherit the project and its tier.
Which Gemini models currently show free pricing?
Google's Gemini Developer API pricing page is the authoritative place to answer this question. On the date above, it shows free input and output for several current text models, including:
gemini-3.6-flashgemini-3.5-flashgemini-3.5-flash-litegemini-3.1-flash-lite
The same page marks some other models unavailable on the free tier. “Gemini has a free tier” therefore does not mean every Gemini model, modality, tool, or preview is free. It also does not establish a request allowance. Pricing answers what one eligible token costs; the rate-limit view answers how much this project may process.
Model rows are volatile. Check the exact ID on the day you build, especially if a tutorial uses a broad label such as “Gemini Pro” or “Flash” without a version. For a first text prototype, gemini-3.6-flash is a practical starting point because it currently has a free row and appears in Google's current getting-started example. That is a starting recommendation, not a promise that it will remain the default.
Your real allowance lives in the project
Google's rate-limit documentation describes three main dimensions:
| Limit | What it measures | Typical failure pattern |
|---|---|---|
| RPM | Requests per minute | A burst or too much concurrency fails quickly |
| TPM | Input tokens per minute | A few large documents can exhaust it |
| RPD | Requests per day | Calls work until the daily pool is spent |
Every request is evaluated against the relevant limits, and exceeding any one can produce a rate-limit error. RPD resets at midnight Pacific Time. Preview and experimental models may have stricter limits.
Most importantly, limits apply per project, not per API key. Creating another key in the same project does not multiply quota. A notebook, cron job, staging deployment, and demo server can quietly compete for the same pool if they share that project.
Google now directs developers to AI Studio to view effective limits and explicitly says published limits do not guarantee actual capacity. That makes a copied “5 RPM / 50 RPD” style table a poor planning tool, even when it was correct for one model and date. Use it as history at most. For an implementation decision, open the project in AI Studio, select the exact model, and record the displayed RPM, TPM, RPD, and tier.
A quick capacity estimate becomes meaningful only after that check. If a document workflow sends 30,000 input tokens per call, divide the displayed TPM by that input size to estimate the maximum token-bound call rate. If a chat endpoint sends tiny prompts to many users, RPM is more likely to be the first constraint. Keep the estimate below the maximum because free capacity can vary.
Make a current first call
Google's current getting-started guide recommends the GA Interactions API through the google-genai SDK. Keep the key in an environment variable:
bashexport GEMINI_API_KEY="YOUR_API_KEY" pip install -U google-genai
Then send a minimal request:
pythonfrom google import genai client = genai.Client() interaction = client.interactions.create( model="gemini-3.6-flash", input="Explain vector databases in three short sentences.", ) print(interaction.output_text)
After it succeeds, inspect AI Studio again. Verify that usage landed on the project you expected and that the exact model still has the limits you planned around. This catches a wrong environment variable, a key from another project, or a stale model ID before those mistakes become application bugs.
Never put a raw Gemini key in a public frontend or commit it to source control. The free tier does not change normal credential security.
Diagnose 429 by the exhausted dimension
429 RESOURCE_EXHAUSTED is a category, not a complete diagnosis.
- A failure immediately after concurrent traffic usually points to RPM. Reduce concurrency and add jittered exponential backoff.
- Failure with a small request count but very large prompts points to TPM. Shorten, chunk, or schedule the input.
- A project that worked for hours and then stops for the rest of the day may have exhausted RPD. Backoff cannot refill a daily bucket.
- Several keys failing together points back to shared project limits.
- A preview model failing while a stable model works can reflect stricter preview capacity.
Retrying makes sense for a short window or transient capacity event. It does not solve an exhausted daily allowance. If you need a deeper error decision tree, use the focused Gemini API 429, 400, and 500 guide.
Free service has a data contract
Google's Gemini API Additional Terms say content submitted to non-paid services, along with generated responses, may be used to provide, improve, and develop Google products and machine-learning technologies. Human reviewers may process that material, and Google tells users not to submit sensitive, confidential, or personal information to non-paid services.
That makes unpaid service a poor fit for private source code, customer records, contracts, health or financial documents, and internal material your organization does not allow to be used for model improvement. Paid-service prompts and responses are not used to improve Google's products under the paid-service terms, although normal security, retention, and compliance review still applies.
Region can also change the answer. Google's availability page lists supported locations. The current terms additionally require paid services when an API client is made available to users in the EEA, Switzerland, or the UK. An individual experiment and a public product therefore may reach different decisions even in the same country.
When paying is the cheaper engineering decision
Stay unpaid when the workload is low-volume, non-sensitive, easy to queue, and inexpensive to interrupt. Enable billing—or choose a different route—when any of these becomes true:
- real users depend on predictable availability;
- free-tier capacity forces operational workarounds;
- prompts contain data that is unsuitable for non-paid terms;
- end-user geography requires paid service;
- a paid-only model or capability is necessary;
- the engineering time spent protecting a tiny quota exceeds the API cost.
Google's billing guide says moving from free to paid requires linking a Cloud Billing account. The current prepaid setup requires at least USD 10 in paid credit. API keys inherit the project and billing state; they do not have independent billing switches.
Do not confuse this with Google's general USD 300 Cloud welcome credit. Google says welcome credit for billing accounts opened after March 2, 2026 cannot pay for Gemini API or AI Studio usage. If you enable billing, set project spending controls, monitor the prepaid balance, and keep tracking RPM, TPM, and RPD rather than assuming payment removes every limit.
For a separate view of current token prices, see the Gemini API token pricing guide.
The durable answer
Gemini API free access remains useful in 2026, especially for learning, prompt validation, hackathons, and low-traffic prototypes. The durable way to use it is not to memorize a quota table. Confirm the region and data contract, pick a model with a current free price row, read the project's live limits in AI Studio, make a minimal call, and observe usage.
Once privacy, reliability, geography, or engineering effort becomes the real constraint, the free phase has done its job. Treat paid service as a production decision, not as a failure to optimize the free tier.
